DEV Community

Fluxenith
Fluxenith

Posted on Fully Autonomous

I built CodexGPT Bridge to use ChatGPT Web models inside Codex (open source)

#opensource #ai #showdev #devtools

Hi DEV! 

I've been experimenting with a way to keep Codex's coding workflow while using the ChatGPT Web models available through my account.

Switching between the two manually gets tedious: I have to move context across windows, paste responses back into the coding tool, and then check whether a suggested fix actually passes tests. So I built CodexGPT Bridge.

It's an MIT-licensed, open-source desktop project for Windows, macOS, and Linux.

GitHub: https://github.com/Fluxenith/codexgpt-bridge

What it does

The idea is to let each side handle what it is good at:

  • ChatGPT Web provides model responses from a signed-in session.
  • Codex keeps control of the local workspace, file changes, commands, tests, and tool approvals.
  • CodexGPT Bridge connects the two so the task can move between reasoning and execution.

I wanted this to feel like a normal coding session rather than a copy-and-paste conversation.

The app also supports separate tasks, image input, progress updates, and cancellation, which matter once you move beyond a single prompt.

A real example

I recorded a 46-second demo showing a bug being reproduced, a fix being applied, and the tests passing afterward. It's included in the repository README:

See the demo and source code on GitHub

For me, that end-to-end loop is the interesting part: the model can help think through the problem while Codex still performs and verifies the actual development work.

How to try it

  1. Open the GitHub repository and follow the README setup instructions.
  2. Set up the desktop bridge and the required Codex/MCP integration and permissions.
  3. Sign in to your ChatGPT Web account.
  4. Start with a small coding task, review any requested tool approvals, and inspect the result.

A few practical notes:

  • Model availability and usage limits depend on your ChatGPT account.
  • The full integration needs the documented MCP/plugin setup and permissions.
  • The macOS builds are not yet notarized.
  • This is an independent community project and is not affiliated with OpenAI.

I'd love feedback

The project is still evolving. If you try it, I'd especially like to hear about setup friction, longer-running tasks, and cases where the handoff between model reasoning and local tools feels awkward.

Issues, suggestions, and contributions are welcome in the repository.

Would you use something like this in your coding workflow? What would you want it to handle better?

Top comments (0)